Passage Ranking

Traditional ranking scored whole documents, which buried good answers inside long pages on broader topics. Passage ranking changed that: systems segment pages into passages — usually bounded by headings — and score each independently against the query. A single excellent section deep in a long guide can now surface for a specific question even if the page overall targets something else.

AI search pipelines take this further. Retrieval-augmented engines chunk documents into passages, embed them as vectors, retrieve the closest matches to a query, and often apply reranking before the model synthesizes an answer. The unit of competition is the passage, not the page.

When ChatGPT search or Perplexity cites a source, it is almost always lifting a specific passage that answered the question on its own. This is the central structural insight of generative engine optimization: a page is not one asset but a collection of potential answers, and each section either survives extraction as a complete, useful passage or it does not.

Passages that win citations share a pattern: a descriptive heading matching the question, a direct answer in the first sentence, supporting specifics — numbers, names, steps — and no dependence on surrounding context. Pronouns referring to earlier sections, or answers split across distant paragraphs, break extraction.

Write each section as a standalone answer: question-shaped heading, conclusion first, evidence after. Keep one idea per section and use semantic HTML so chunkers find the boundaries you intended. Front-load a 40-60 word summary at the top of important pages, since that block often becomes the cited passage. Geonimo's page optimization analysis scores content at this passage level, flagging sections that depend on context or bury their answers — the exact failures that keep otherwise strong pages out of AI citations.